Smart robots are already in many households, autonomous vehicles will drive on our roads presumably in the next decade and Artificial Intelligence (AI) Systems are helping more and more to diagnose and treat diseases. Intelligent systems are ubiquitous today and it is just the beginning. The socially responsible and beneficial development of AI has never been more important than nowadays, where problems range from estimating causal effects of treatments for diseases to removing unwanted confounders and racial discrimination in automatic decision making. 

In these cases if we just look at data we often find spurious correlations, leading to potentially catastrophic results. In order to draw causal conclusions from images, we need to discover high-level causal variables from low-level observations like pixel values. My research focuses on the intersection of causal inference and deep learning with a particular focus on learning causal representations. In the longterm, my research goal is to design machines that can extrapolate experience across environments and tasks.

In recent years, our efforts have been focused on creating benchmarks for evaluating causal models for different downstream tasks where we have access to the ground truth. For instance, in robotics, we use knowledge about their design for testing and improving the generalization capabilities of different algorithms in real-world settings. Using these robotic benchmarks, we hosted a robotics competition in the cloud in 2020 real-robot-challenge.com, which is a small step towards democratizing robotics. A key motivation for organizing these challenges is that currently, only very academic and industrial labs have access to the software and hardware setups needed to perform machine and reinforcement learning research directly on real-world systems. The challenge made the research in our field more inclusive by reducing the entry barriers significantly and we were happy to receive submissions from across the globe.

Beyond robotics, I am interested in AI applications, for example in healthcare. We analyzed 70k patients across 70 hospitals in the US and Europe for real-time COVID-19 related mortality prediction https://www.nature.com/articles/s41467-020-20816-7 or created enabling tools like web platforms for high-throughput analysis of sleep patterns, which are used daily by academic and industrial researchers worldwide.